Optical emission spectroscopy (OES) is an important technique for plasma diagnostics. However, random deviations in emission spectra measurements are inevitable due to instrumental imperfections and other interferences. In scenarios requiring high temporal resolution measurements, where repeated measurements are impractical, these random errors pose significant challenges for accurate plasma diagnostics. This work introduces a novel OES method that utilizes a neural network model to suppress random deviations in emission spectra measurements. The dataset for training neural network is generated using a comprehensive collisional-radiative model combined with an instrument disturbance model. The novel method is demonstrated on a microwave electron-cyclotron-resonance discharge chamber. The results show that the novel method reduces the random deviation in electron temperature and density to less than 3%, which represents a significant improvement over traditional methods. Additionally, the new OES method offers enhanced timeliness, making it particularly promising for online monitoring applications.